Goldman Sachs warns next-gen AI servers could trigger localized blackouts

AI data centers are no longer just a tech story. They are becoming an energy crisis in slow motion, and Goldman Sachs just put a timestamp on it.

The bank’s commodity and energy analysts, including Samantha Dart, co-head of global commodities research, are warning that connecting next-generation AI servers directly to the US power grid could cause localized blackouts, particularly in the dense clusters of counties where most data center capacity is concentrated.

The power math is staggering

Here is the core problem: next-gen AI server racks expected to deploy around 2027 could each consume more than 500 to 600 kilowatts. That is roughly 50 times the draw of older rack designs.

Goldman’s analysts project that US data center power demand could reach approximately 108 GW by 2030.

The geographic concentration makes this worse. According to Goldman’s research, roughly 72% of US data center capacity sits in just 1% of counties.

Reserve margins, the buffer between peak demand and total available capacity, are projected to dip below 15% across most US grids if capacity additions do not accelerate. Grid operators generally consider anything under 15% a warning zone. Below that threshold, even routine demand spikes during summer heat waves can tip a region toward rolling outages.

The infrastructure backlog problem

Building more power plants sounds like the obvious fix. The less obvious part is that getting new generation connected to the grid can take more than seven years in some regions due to interconnection queue backlogs.

Goldman’s report notes that equipment shortages, particularly transformers, compound the delay problem. Transformer lead times have stretched dramatically in recent years, and the specialized high-capacity units needed for data center campuses are among the hardest to source quickly.

Community opposition adds another variable. Local residents near proposed substations and transmission corridors have become more organized in pushing back on large infrastructure projects, adding regulatory and legal friction to an already slow process.

The result is a growing push toward what the industry calls behind-the-meter power. Rather than drawing from the public grid, data centers build or contract for dedicated generation capacity that bypasses the interconnection queue entirely. Goldman’s analysts suggest that as much as one-third of future AI data center capacity could end up operating as effectively islanded power systems.

What this means for energy markets and the AI race

Goldman’s analysts point to natural gas peaker plants and renewable energy as the most likely beneficiaries of accelerated investment. Peakers are fast-response generation assets that can spin up quickly during demand spikes, making them particularly valuable in a grid environment where reserve margins are thinning.

Utilities in regions with more agile infrastructure, faster permitting, existing transmission headroom, and diversified generation mixes, stand to gain a disproportionate share of future data center investment. Goldman’s research frames this as a potential competitive edge in the broader AI economy.

Goldman’s warning, at its core, is that the energy system was not built for this rate of change, and the gap between AI infrastructure ambitions and grid reality is widening faster than utilities and regulators are currently moving to close it.

Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.